Modular Software Performance Monitoring

Daniele Francesco Kruse – CERN – PH / SFT

Karol Kruzelecki – CERN – PH / LBC

Summary1. Motivation and Goal

2. Introduction• Performance Monitoring• Performance Counters

3. Challenge• Monolithic vs. Modular Monitoring• Modularity in CMSSW, Gaudi and Geant4

4. Solution• Analysis Overview• What do we monitor?

5. Simple example of successful usage

6. Conclusions2

Motivation and Goal• HEP software is huge and complex and

is developed by a multitude of programmers often unaware of performance issues

• The software produced is suboptimal in terms of efficiency and speed

• Unfortunately CPU speed is not likely to be increased in the near future as we were used to in the past

• The goal is then to find an effective method to improve SW through monitoring and optimization

• Better performance (more throughput) means savings both in hardware and power needed

3

Performance Monitoring

DEF : The action of collecting information related

to how an application or system performs

HOW : Obtaining micro-architectural level information

from hardware performance counters

WHY : To identify bottlenecks, and possibly remove them

in order to improve application performance4

Performance Counters• All recent processor architectures

include a processor–specific PMU

• The Performance Monitoring Unit contains several performance counters

• Performance counters are able to count micro-architectural events from many hardware sources (cpu pipeline, caches, bus, etc…)

• We focus on the two main Intel® cpu families currently on the market: Core and Nehalem

• Nehalem processors feature 4 programmable counters while Core processors have 2 programmable counters

5

Monolithic vs. Modular Monitoring

Monolithic“Black Box”

1 RUN 1 set of results

VS.

6

Res. A1

Res. A2

Res. A3

Res.

AN

Res. A

AVG.

Res. C1

Res. C2

Res. C3

Res.

CN

Res. C

AVG.

Res. B1

Res. B2

Res. B3

Res.

BN

Res. B

AVG.

Module C

Module B

Module A

Event 1

Event 2

Event 3

Event N

1 RUN

set of results for A

set of results for B

set of results for C

For each CPU performanceevent that we monitor...

Monolithic vs. Modular Monitoring

• When we face large and complex software monolithic analysis becomes less useful

• “Traditional” monitoring tools (using performance counters) are monolithic. Examples: PTU and pfmon

• Even sampling over symbols (functions) is not enough for code division. Solution: modular monitoring!

• Code instrumentation (minimal in HEP software) and ad-hoc interface to the monitoring tool needed

• Advantage: narrowing down the possible location of performance problems leads to easier optimization

7

Modularity in HEP SW• CMSSW code is organized into modules

that are sequencially executed during each event processed, and it provides hooks to execute user defined actions at the beginning and at the end of modules

• Hooks is what we use to start and stop the monitoring process and to collect results for each module

• More on CMSSW performance in Matti Kortelainen’s talk

• Gaudi provides a similar mechanism to instrument its code (modules are called algorithms)

• Geant4 is handled differently: binning into triples<particle type, energy range, physical

volume>

8

Analysis flow-graph

Analysis Configurati

on

Start

PerformanceData Taking

Program Run

Performance Data Output

PerformanceData

Analysis

Browsable HTML results

End

9

CMSSWor

Gaudior

Geant4

What and why do we monitor?

Total Cycles (Application total execution time)

Issuing μops Not Issuing μops

Stalled(no work)

Not retiring μops

(useless work)

Retiring μops

(useful work)

Ifetch

misses

Load Stalls

FP Exceptions

Branches

Divs & Sqrts

10

Example of usage – LHCb (Gaudi)

• Tested on Brunel v37r7

GaudiRun <options>

GaudiProfiler <options>• GaudiProfiler – python script handling sequential run of application for all the necessary counters and postprocessing

No code instrumentation needed

11

Choice of an alghorithm• Objective: to reduce the number of Total

cycles (execution time) of one algorithm• As a simple example we choose to focus

on reducing the cycles in which instructions were retired

• How: This is done by reducing the number of Instructions Retired – a very stable and reliable counter

12

Symbols – Looking deeper in

• After choosing Algorithm, in the detailed symbol view .cpp file and function

• Inlined functions are not shown, they are counted in the “parent” functions

13

Detailed profiling of one function

• We modify the body of function adding start() and stop() commands for profiler

• Results is shown after the run of application is over

14

Improving small parts of code

• Optimization procedure loop:1. modify

code2. compile3. profile

• Compare average count of Instructions Retired

-6.5 %

15

Re-run after changes

• Even small changes are visible in Instructions Retired

• Total Cycles decreased

– it is faster

• ~6.5% improvement in one function gave ~2% improvement in the algorithm

16

Conclusions• We implemented a modular ad-hoc

performance counters-based monitoring tool for three major HEP frameworks: CMSSW, Gaudi and Geant4

• This tool is supposed to help developers optimizing existing code to improve its performance without the need for code instrumentation

• The tool has been successfully used to optimize code in Gaudi and has shown the potential to be used for other applications as well

• GaudiProfiling package will be available in the next release of Gaudi

17

Thank you, Questions ?

backup slides

19

BACKUP: The 4-way Performance Monitoring

1. OverallAnalysis

2. Symbol LevelAnalysis

3. Module LevelAnalysis

4. Modular SymbolLevel Analysis

Overall (pfmon)

Modular

Sampling

Counting

20

BACKUP: Core and Nehalem PMUs - Overview

Intel Core Microarchitecture PMU• 3 fixed counters

(INSTRUCTIONS_RETIRED, UNHALTED_CORE_CYCLES, UNHALTED_REFERENCE_CYCLES)

• 2 programmable counters

Intel Nehalem Microarchitecture PMU• 3 fixed core-counters

(INSTRUCTIONS_RETIRED, UNHALTED_CORE_CYCLES, UNHALTED_REFERENCE_CYCLES)

• 4 programmable core-counters• 1 fixed uncore-counter (UNCORE_CLOCK_CYCLES)

• 8 programmable uncore-counters21

BACKUP: Perfmon2• A generic API to access the PMU (libpfm)• Developed by Stéphane Eranian• Portable across all new processor micro-

architectures• Supports system-wide and per-thread

monitoring• Supports counting and sampling

CPU Hardware

Linux KernelGeneric Perfmon

Architectural Perfmon

PMU

User spacePfmon Other libpfm-based Apps

libpfm

BACKUP: Nehalem : Overview of the architecture

Core 0 Core 1 Core 2 Core 3

Level 3 Cacheshared, writeback (lazy write) and inclusive (contains L1

& L2 of each core)Integrated Memory

ControllerLink to local memory

(DDR3)

Quick Path InterconnectLink to I/O hub

(& to other processors, if present)

L1D Cache

L1I Cache

L2 Cache

TLBs

writeback

writeback

unified, writeback, not-inclusive

DTLB0 & ITLB (1st Level), STLB (unified 2nd Level)

23

BACKUP: μops flow in Nehalem pipeline

Instruction Fetch& Branch Prediction

Unit

Decoder

Retirement&

Writeback

Re-OrderBuffer

ResourceAllocator

Execution

Units

Reservation

Station

UOPS_ISSUED

UOPS_RETIRED

UOPS_EXECUTED

• We are mainly interested in UOPS_EXECUTED (dispatched) and UOPS_RETIRED (the useful ones).

• Mispredicted UOPS_ISSUED may be eliminated before being executed. 24

BACKUP: Cycle Accounting Analysis

Total Cycles (Application total execution time)

Issuing μops Not Issuing μops

Stalled(no work)

Not retiring μops

(useless work)

Retiring μops

(useful work)

Store-FwdL2 miss

L2 hit

LCP

L1 TLB miss

BACKUP: New analysis methodology for NehalemBASIC STATS: Total Cycles, Instructions Retired, CPI;

IMPROVEMENT OPPORTUNITY: iMargin, iFactor;

BASIC STALL STATS: Stalled Cycles, % of Total Cycles, Total Counted Stalled Cycles;

INSTRUCTION USEFUL INFO: Instruction Starvation, # of Instructions per Call;

FLOATING POINT EXCEPTIONS: % of Total Cycles spent handling FP exceptions;

LOAD OPS STALLS: L2 Hit, L3 Unshared Hit, L2 Other Core Hit, L2 Other Core Hit Modified, L3 Miss -> Local DRAM Hit, L3 Miss -> Remote DRAM Hit, L3 Miss -> Remote Cache Hit;

DTLB MISSES: L1 DTLB Miss Impact, L1 DTLB Miss % of Load Stalls;

DIVISION & SQUAREROOT STALLS: Cycles spent during DIV & SQRT Ops;

L2 IFETCH MISSES: Total L2 IFETCH misses, IFETCHes served by Local DRAM, IFETCHes served by L3 (Modified), IFETCHes served by L3 (Clean Snoop), IFETCHes served by Remote L2, IFETCHes served by Remote DRAM, IFETCHes served by L3 (No Snoop);

BRANCHES, CALLS & RETS: Total Branch Instructions Executed, % of Mispredicted Branches, Direct Near Calls, Indirect Near Calls, Indirect Near Non-Calls, All Near Calls, All Non Calls, All Returns, Conditionals;

ITLB MISSES: L1 ITLB Miss Impact, ITLB Miss Rate;

INSTRUCTION STATS: Branches, Loads, Stores, Other, Packed UOPS; 26

BACKUP: PfmCodeAnalyser, fast code monitoring• Unreasonable (and useless) to run a

complete analysis for every change in code

• Often interested in only small part of code and in one single event

• Solution: a fast, precise and light “singleton” class called PfmCodeAnalyser

• How to use it:#include<PfmCodeAnalyser.h>

PfmCodeAnalyser::Instance(“INSTRUCTIONS_RETIRED”).start();

//code to monitor

PfmCodeAnalyser::Instance().stop(); 27

BACKUP: PfmCodeAnalyser, fast code monitoringPfmCodeAnalyser::Instance("INSTRUCTIONS_RETIRED", 0, 0, "UNHALTED_CORE_CYCLES", 0, 0, "ARITH:CYCLES_DIV_BUSY", 0, 0, "UOPS_RETIRED:ANY", 0, 0).start();

Event: INSTRUCTIONS_RETIRED Total count:105000018525Number of counts:10 Average count:10500001852.5

Event: UNHALTED_CORE_CYCLES Total count:56009070544Number of counts:10 Average count:5600907054.4

Event: ARITH:CYCLES_DIV_BUSY Total count:28000202972Number of counts:10 Average count:2800020297.2

Event: UOPS_RETIRED:ANY Total count:138003585913Number of counts:10 Average count:13800358591.3 28

BACKUP: What can we do with counters?

Question:

Is all this useful?

Answer:We don’t

know,but we shall

see• Lack of papers and literature about the

subject

• An empirical study is underway to find out:1. A relationship between counter results and coding

practices2. A practical procedure to use counter results to

optimize a program

• A procedure has already been developed and will be tested

• The trial study will be conducted on Gaudi together with Karol Kruzelecki (PH-LBC group)

29

BACKUP: The 3-step optimization procedure• We start from counter results and choose one

algorithm to work on using the Improvement Margin and the iFactor.

• We then apply the following procedure:MyAlg : Total Cycles: 1000 , Total Instructions: 1000 → CPI: 1.00

MyAlg : Total Cycles: 300 , Total Instructions: 250 → CPI: 1.20

MyAlg : Total Cycles: 250 , Total Instructions: 250 → CPI: 1.00

MyAlg : Total Cycles: 230 , Total Instructions: 250 → CPI: 0.92

1. Change to a more efficient algorithm and vectorize it

2. Remove stall sources (L1 & L2 misses, store-fwd, etc..)

3. Remove misprediction sources (branches, calls, etc..)

30

BACKUP: Overall Analysis

• Uses Pfmon and it is based on the Cycle Accounting Analysis

• Good for showing overall performance and for checking improvements

• Good for identifying general software problems

• Good for comparing different versions of the code

• NOT enough for• finding inefficient parts of the software• finding bad programming practices

BACKUP: Overall Analysis

BACKUP: Symbol Level Analysis

• Uses sampling capabilities of pfmon

• Good for identifying general bad programming practices

• Can identify problems of functions which are frequently used

• Shows functions that use most of the execution cycles and functions that spend a lot of time doing nothing (stalling)

• NOT good for finding specific problems in the code

Stalled Cycles

BACKUP: Symbol Level Analysis

Total Cyclescounts %self symbol 54894 3.79% _int_malloc 50972 3.52% __GI___libc_malloc 41321 2.85% __cfree 36294 2.51% ROOT::Math::SMatrix::operator= 31100 2.15% __ieee754_exp 25636 1.77% ROOT::Math::SMatrix::operator= 24833 1.72% do_lookup_x 23206 1.60% ROOT::Math::SMatrix::operator= 22970 1.59% __ieee754_log 21741 1.50% __atan2 20467 1.41% ROOT::Math::SMatrix::operator= 19922 1.38% _int_free 18354 1.27% G__defined_typename 16026 1.11% strcmp 15979 1.10% TList::FindLink 14601 1.01% G__defined_tagname

counts %self symbol 24955 5.09% _int_malloc 19797 4.04% do_lookup_x 19084 3.89% __GI___libc_malloc 14282 2.91% __ieee754_exp 13564 2.77% strcmp 13065 2.66% __cfree 9927 2.02% __atan2 8998 1.83% __ieee754_log 7666 1.56% TList::FindLink 7575 1.54% _int_free 5392 1.10% std::basic_string::find 4911 1.00% computeFullJacobian 4410 0.90% malloc_consolidate 4285 0.87% operator new 4104 0.84% ROOT::Math::SMatrix::operator= 3949 0.81% 33.84% makeAtomStep

BACKUP: Module Level Analysis• Uses the Perfmon2 interface (libpfm)

directly

• Analyses each CMSSW module separately

• Allows the identification of “troubled” modules through a sortable HTML table

• Gives instruction statistics and produces detailed graphs to make analysis easier

• It requires 21 identical cmsRun’s (no multiple sets of events are used → more accurate results), but it can be parallelized so (using 7 cores): time = ~3 runs

• Code outside modules is not monitored (framework)

DEMO

BACKUP: Module Level Analysis - Results Snapshot

BACKUP: Single Module Graphs

37

BACKUP: Modular Symbol Level Analysis• Uses the Perfmon2 interface (libpfm)

directly and analyses each CMSSW module separately

• Sampling periods are specific to each event in order to have reasonable measurements

• The list of modules is a HTML table sortable by number of samples of UNHALTED_CORE_CYCLES

• For each module the complete set of usual events (Cycle Accounting Analysis & others) is sampled

• Results of each module are presented in separate HTML pages in tables sorted by decreasing sample count

38

BACKUP: The List of Modules

39

BACKUP: Table Example of a Module

40

Version with

graphs

BACKUP: Structure and libraries

Analysis Configurati

on

Start

PerformanceData Taking

Program Run

Performance Data Output

PerformanceData

Analysis

Browsable HTML results

End

libpng

zliblibSDL

libSDL_ttflibpfm

zlib

41

BACKUP: The Sampling Process

Algorithmexecutes its code

Sampling starts

Buffer full?

Buffer flushescontent to memory

Algorithm stopsexecuting its code

Sampling stops

Algorithmfinished?

End

Start

Yes

No

Yes

No

For each algorithm...Start

Samplerinitialization

Samplertermination

Samples dumpedto output files

EndBuffer flushes

content to memory

Sampling stops

42

BACKUP: Module Level Analysis Results Snapshot

BACKUP: Improvement Margin and iFactor

iMargin (CPI reduction effects)

data: cur_CPI, exp_CPI, local_cyc_bef, glob_cyc_before.• loc_imp_ratio = cur_CPI/exp_CPI• loc_cyc_after = local_cyc_bef/ loc_imp_ratio • glob_cyc_after = glob_cyc_before – loc_cyc_before +

loc_cyc_after• improvement_margin = 100 –

(glob_cyc_after/glob_cyc_before) * 100iFactor (Improvability Factor)

data: simd_perc, missp_ratio, stalled_cycles.• simd_factor = 1 – normalized(simd_perc)• missp_factor = normalized(missp_ratio)• stall_factor = normalized(stalled_cycles)• iFactor = stall_factor * (simd_factor + missp_factor +

stall_factor) 44

BACKUP: “ Vertical ” vs. “ Horizontal ” cut• For Gaudi and CMSSW we used a “

horizontal ” cut

• Geant4 doesn’t provide hooks for any horizontal cut

• Modular analysis through User Actions

• “ Time division ” instead of “ Code division ”

• Useful or not? maybe... taking particles, energies and volumes into consideration

• Moreover modular symbol analysis still provides “ Code division ”

BACKUP: Choice of granularity• Different levels of granularity were

considered (run, event, track and step) as each offered User Actions

• Step-level granularity was the final winner

• At each step the particle, its energy (at the beginning) and the physical volume that it is running through are used

• Interesting volumes (at any level in the geometry tree) are given through an input file and used in the results view

• Other volumes are labeled as “OTHER”

• Results are browsable by any of the above variables

BACKUP: “ Total ” vs. “ Average ” count• CMSSW and Gaudi used average counts of

performance events

• All modules were “used” the same number of times during a single execution

• No longer true in Geant4 steps since “modules” here are a combination of physics variables

• Therefore we chose to display total counts of all performance counters

• Exception: for the number of UNHALTED_CORE_CYCLES we provide both average and total counts

BACKUP: (1/3) How to use it?

• Unpack the following archive in your application directory:

http://dkruse.web.cern.ch/dkruse/G4_pfm.tar.gz

• Add the following lines to your GNUmakefile (to link libpfm):

CPPFLAGS += -I/usr/include/perfmonEXTRALIBS += -lpfm -ldl

-L/afs/cern.ch/sw/lcg/external/libunwind/0.99/x86_64-slc5-gcc43-opt/lib -lunwind

EXTRALIBSSOURCEDIRS += /afs/cern.ch/sw/lcg/external/libunwind/0.99/x86_64-slc5-gcc43-opt

• Edit the “RUN_CONFIG” attribute in the pfm_config_arch.xml file inserting the normal run command. Example:

RUN_CONFIG=“~/geant4/bin/Linux-g++/full_cms bench10.g4”

• Edit the pvs.txt file inserting the interesting physical volumes:

CALO MUONVCAL BEAMTRAK

48

BACKUP: (2/3) How to use it?

• Add the following lines to your main() before runManager->Initialize():

#include "PfmSteppingAction.hh”...int base_arg_no = 2; //number of standard arguments including executable namerunManager->SetUserAction(new PfmSteppingAction(argv[base_arg_no], argv[base_arg_no+1], atoi(argv[base_arg_no+2]), (unsigned int)atoi(argv[base_arg_no+3]), (unsigned int)atoi(argv[base_arg_no+4]), argv[base_arg_no+5], atoi(argv[base_arg_no+6]), (unsigned int)atoi(argv[base_arg_no+7]), (unsigned int)atoi(argv[base_arg_no+8]), argv[base_arg_no+9], atoi(argv[base_arg_no+10]), (unsigned int)atoi(argv[base_arg_no+11]),(unsigned int)atoi(argv[base_arg_no+12]), argv[base_arg_no+13], atoi(argv[base_arg_no+14]), (unsigned int)atoi(argv[base_arg_no+15]),(unsigned int)atoi(argv[base_arg_no+16]),(unsigned int)atoi(argv[base_arg_no+17]), argv[base_arg_no+18], (bool)atoi(argv[base_arg_no+19])) );

• Compile and link:gmake g++ -Wall -o create create_config_files_from_xml.cpp -lxerces-c g++ -Wall -lz -o analyse pfm_gen_analysis.cpp

49

BACKUP: (3/3) How to use it?

• Create results directory:mkdir results

• Create python run script “G4perfmon_runs.py”:./create pfm_config_nehalem.xml

• Run the application with the perfmon monitor:python G4perfmon_runs.py &

• Analyse the results (optionally generating csv file):

./analyse results/ --caa [--csv]

• Check your results using your favourite browser:

firefox results/HTML/index.html50

BACKUP: XML configuration file

<?xml version="1.0" ?><PFM_CONFIG> <PROPERTIES NAME="GeneralAnalysis" RUN_CONFIG="hlt_HLT.py" OUTPUT_DIR="results/" /> <CONFIG START_AT_EVENT="4" PARALLEL="0" /> <EVENTS> <EVENT_SET> <EVENT NAME="BR_INST_RETIRED:ALL_BRANCHES" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="ILD_STALL:ANY" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="MEM_INST_RETIRED:LOADS" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="MEM_INST_RETIRED:STORES" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> </EVENT_SET> <EVENT_SET> <EVENT NAME="INST_RETIRED:ANY_P" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="ITLB_MISS_RETIRED" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="MEM_LOAD_RETIRED:DTLB_MISS" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> <EVENT NAME="MEM_LOAD_RETIRED:L2_HIT" CMASK="0" INVMASK="0" SMPL_PERIOD="0" /> </EVENT_SET> <EVENT_SET> <EVENT NAME="ARITH:CYCLES_DIV_BUSY" CMASK="0" INVMASK="0" SMPL_PERIOD="1000" /> </EVENT_SET> </EVENTS></PFM_CONFIG>

51